AI and the 2048 Game

A multi-part series exploring how artificial intelligence can be applied to the 2048 game. Covers human strategies, expectimax, heuristics, optimization, and self-learning techniques, with live browser demos.

Illustration: Automatic 2048-Solving Algorithm: From Human Strategy

Automatic 2048-Solving Algorithm: From Human Strategy to Expectimax AI

20/11/2025 0 34 6 min read

Demo

The game 2048 looks very simple, but if the goal is to build a bot that can automatically reach 2048, 4096, or even 8192 consistently, then this is no longer just a casual puzzle game...

Illustration: Self-learning 2048 bot: tuning the evaluation

Self-learning 2048 bot: tuning the evaluation function with self-play

20/11/2025 0 22 5 min read

Demo

In Part 1, we built a 2048 bot using expectimax plus a hand-crafted evaluation function: more empty cells, smoothness, monotonicity, big tile in the corner, etc. The bot plays decently, but there is still a...

Illustration: AI & 2048: Learning the Evaluation

AI & 2048: Learning the Evaluation Function with TensorFlow.js

20/11/2025 0 37 4 min read

Demo

In Part 1, we built a 2048 bot using expectimax plus a hand-crafted evaluation function. In Part 2, we let the bot self-tune that evaluation function using simple self-play hill-climbing. In this Part 3, we...